ChinAI #346: Reputation Collectives - how international industry associations have helped raise China's safety standards in high-risk technologies

My latest article on how international private governance could play a role in AI governance

Feb 09, 2026

Greetings from a world where…

curling is captivating even if one does not know the rules

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Reputation Collectives: how international industry associations influence China’s safety standards in high-risk technologies

Many policymakers and researchers see China as the most likely source of an AI accident. In a 2023 Foreign Affairs article, Bill Drexel and Hannah Kelley declared, “Due to Beijing’s lax approach toward technological hazards and its chronic mismanagement of crises, the danger of AI accidents is most severe in China.” In some ways, this is an incredibly bold claim to make, given the consequences — U.S. policymakers and AI lab leaders exploit this assumption to justify their own inaction on AI safety.

In another sense, this claim is not very bold at all. After all, it is in line with conventional wisdom and empirical evidence that technological accident risks are high in authoritarian regimes due to limited transparency and civil society guardrails. See: Chernobyl.

Yet, China has achieved remarkable safety gains in certain high-risk technological domains, including civil aviation and nuclear power. How?

Co-authored with GW PhD student Dennis Li, my latest article in Review of International Political Economy presents an unexpected answer: in industries where one firm’s accident damages the reputation of all others, international industry associations can help improve safety standards in emerging economies. In these high-risk domains, firms must protect the industry’s reputation as a shared resource:

For industries in which an accident in one company damages the reputation of all others, industry reputation functions as a common-pool resource that motivates firms to collectively monitor its consumption and discourage misuse (i.e., track and deter any reputation-depleting actions) (Barnett & King, 2008, p. 1152; Ostrom, 1990). For example, after the Three Mile Island accident, Bill Lee, president of a major US utility company, spearheaded the creation of the Institute of Nuclear Power Operations (INPO), which is often held up as an exemplary model for industry self-regulation. In a speech after the accident, Lee aptly captured the notion of an industry’s collective reputation when he stated that all nuclear power plants were ‘hostages of each other’ (Cantelon, 2016, p. viii, emphasis added).

We label this the “reputation collectives” mechanism of international private regulation, differentiating it from a more familiar and well-studied mechanism of “certification clubs” (think: fair trade coffee). For instance, when it comes to peer pressure, certification clubs rely on public naming and shaming: completely transparency allows NGOs and the public to shower praise on safety leaders and heap scorn on laggards. In contrast, reputation collectives employ internal benchmarking: members share safety results and conduct peer reviews, but this information is not disclosed outside the group in a way that could single out a particular firm (see figure below).

An earlier version of this paper was titled “Industrialists Anonymous”, as a nod to Alcoholics Anonymous, another peer group that regulates behavior by encouraging candid disclosures within a confidential setting.

Drawing on expert interviews and Chinese-language sources, we examine interactions between international industry associations and Chinese firms in three high-risk technological domains: nuclear power (1987-2016), civil aviation (1990-2008), and chemicals (2002-2021). In each of these cases, the relevant industry association — the World Association of Nuclear Operators (WANO), the International Air Transport Association (IATA), and the International Council of Chemical Associations, respectively — shaped China's safety improvements through the reputation collectives pathway. They treated industry reputation as a communal good, guarded information on member performance from external stakeholders, and subsidized weak-link firms to improve their safety practices.

When I started this project many years ago, I was expecting to trace interactions between the IAEA or ICAO (the international public regulators) and Chinese authorities. However, when I dug more into these cases, I grew more curious about the role of international industry associations WANO and IATA. I want to share two snippets from the empirical cases that demonstrate how Chinese interactions with the international private regulators sometimes raised standards beyond what the public regulators could implement. First, consider WANO (rhymes with Bono):

Next, here’s a snippet from our IATA case study:

We conclude by considering the implications for the governance of AI, in which voluntary industry commitments are becoming more prevalent: “Our paper suggests that an important variable for the design of these initiatives is whether the AI industry develops a collective safety reputation. If it does, then the effectiveness of global private governance will rest on the features of reputation collectives: low entry requirements in pursuit of universal membership, avoidance of public naming-and-shaming, and reliance on socialization and peer-to-peer learning to improve the safety performance of laggards.”

What’s the closest thing we have to an industry association dedicated to AI safety? It’s probably the Frontier Model Forum, but it only includes six U.S. companies and weirdly frames its mission to be “focused on addressing significant risks to public safety and national security (emphasis mine).” This national approach will likely prove inadequate, if our paper’s historical insights are taken seriously.

Full article available open-access: Reputation collectives - how international industry associations influence China’s safety standards in high-risk technologies

ChinAI Links (Four to Forward)

Two follow-ups on industry-led efforts in AI Safety/Security Governance Report

A few weeks ago, I flagged the China AI Industry Alliance’s “AI Security and Safety Commitments.” Currently, 22 have signed on to these commitments, and 18 have disclosed their security and safety practices as part of the alliance’s voluntary disclosure initiative. Two follow-ups on this front:

AI governance researcher Julia Chen shared with me an article indicating that the sharing of company disclosures in July 2025 refers to a list of 43 examples of typical practices on the Security and Safety Commitments website. These are not attributed to particular signatories.
- *My view: these seem way too vague to be helpful; I wish there was more transparency and attribution to particular firms. From what I’ve seen, the two reports that disclose the most specifics on what Chinese companies’ safety and security practices are: 1) Tencent’s Large Model Safety and Ethics Research Report 2024 (ChinAI translation and coverage); 2) Alibaba’s The Large Language Model Technology development and governance practice report. I’ve translated the first for ChinAI but would welcome help on the latter.
A couple weeks ago, I wrote, “These have been released for over a year, but I don’t think there’s been any English-language coverage of them (and I certainly haven’t picked up on them).” Fortunately, I was wrong! Scott Singer had a great Carnegie analysis on these commitments, comparing them to “ongoing global industry-led efforts to put safeguards in place for frontier AI piloted at last year’s AI Summit in Seoul, known as the Seoul Commitments.”

Should-read: China’s AI Boyfriend Business Is Taking On a Life of Its Own

For Wired, Johanna Costigan reports on China’s Gen Z women and the booming AI boyfriend business. One platform Xingye (星野) allows users to customize their AI companions.

Should-read: Support for Washington Post international employees

Sadley, the Post laid off some stellar China correspondents along with their local partners; this type of investigative, fact-checked reporting (with editorial guidance) is not easily replaced. Lily Kuo shared this gofundme link: “I was the Washington Post’s China bureau chief until last year. So much our coverage depended on our local researchers. Some left China so that they could do their work more safely. Now their future is uncertain.”

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

ChinAI #345: A Three-way Race for China's AI Super-App

ByteDance vs. Tencent vs. Alibaba

Jeffrey Ding

Feb 02, 2026

Greetings from a world where…

snowcrete is a word?

Feature Translation: Chronicles of the Great AI Battle among ByteDance, Alibaba, and Tencent — A destiny-changing Conflict

Context: DeepSeek shocked the world last January, giving hope to Chinese frontier model start-ups. In the year since, three Chinese tech giants have struck back. This longform, deeply-sourced LatePost [晚点] article chronicles the three-way contest among ByteDance, Tencent, and Alibaba to build the AI Super-App. Below are notes on key passages, grouped by themes of macro-level strategy, talent acquisition, and comparative advantages of the giants vs. start-ups. The full translation is a 8,500-word, 20-pg. Google Doc.

Key Passages: When it comes to China’s AI super-app, everybody is chasing ByteDance’s Doubao. Moonshot AI’s Kimi made waves with its long-prompt processing capabilities in 2024; in early 2025, of course, DeepSeek’s downloads briefly topped the charts; and Alibaba’s Qwen relaunched successfully at the end of the year. Yet, as Honghao Gao reports, “ Doubao consistently remained in first place.”

Last year, ByteDance’s Doubao became the first AI product in China to surpass 100 million daily active users. Doubao’s large model attained an average daily token processing volume of 63 trillion, a 200% increase over the past six months.
In many ways, ByteDance was a sleeping giant. By the time ChatGPT launched in November 2022, Baidu, Huawei, and Alibaba had already released their own LLMs, but ByteDance had not. The next year, ByteDance’s AI efforts stalled. Their AI creation platform Jimeng did not “become the Douyin of the AI era”, as they hoped.
In fact, the breakthrough was an AI assistant named Doubao, which benefited from ByteDance’s Douyin (shortform video platform) and its unique voice call function. Regarding the former, videos of users “creatively training” Doubao’s voice went viral on Douyin, which boasts 800 million daily active users. As for the voice call function, the Doubao team:

“traveled across the country to collect dialects, refining the granularity of accents down to the district and county level within cities…and they conducted a series of stylized dialogue training sessions on the model, giving it a distinct personality.”

ByteDance also invested in its foundation model capabilities. In early February 2025, it hired Yonghui Wu, former Vice President of Google DeepMind, to manage theoretical research on large language models in its Seed team.

Tencent goes all-in to catchup and leapfrog ByteDance. In 2025, it poured more resources into its Yuanbao chatbot, which is based on its Hunyuan large model. The clearest evidence of this shift came in September 2025, when it poached 27-year-old Shunyu Yao, a former OpenAI researcher.

First, let’s back up. Since 2023, Tencent took a cautious approach to LLMs, neither actively stockpiling compute nor making huge talent acquisitions. Its Yuanbao launched a year later than Doubao, and even then, it was mostly positioned as “a product for Tencent to test its Hunyuan large language model technology.” An early Yuanbao team member told LatePost, “Yuanbao was ‘in a state of begging other internal departments for collaboration.’”
After Yao took charge of Tencent’s large model, he called an internal meeting in November 2025, where he showed up wearing shorts and slippers. Gao reports, ‘There are big problems with Hunyuan's evaluation,’ a person present at the meeting quoted Yao as saying, meaning that the model was overly focused on achieving high rankings on leaderboards, incorporating benchmark data into the training set, leading to data contamination. Although the model was good at answering questions, its performance in real-world scenarios was unstable. He hoped that, in the future, the team wouldn't focus on rankings or let the rankings dictate their work.“
Some of the fixes were unsexy affairs, such as addressing data standardization issues. Here’s the relevant LatePost passage: “A source close to Yuanbao told us that TEG’s data cleaning and labeling in small model training were not standardized. Taking the LBS (location-based services) intent recognition model as an example: a user asking ‘What are some good restaurants nearby? [周边有什么好吃的？]’ is a (prompt with) clear intent; while ‘What would you recommend that are good things to eat [有什么好吃的推荐]?’ is a (prompt with) vague intent. If these two types of samples were mixed and labeled together in the training set, the model would easily ‘learn incorrectly,’ leading to an increased error rate.”
In addition to improving its Hunyuan model, Yuanbao also integrated DeepSeek. Tencent finally put in a big order on computing cards to deploy DeepSeek across Tencent products, and Yuanbao climbed into the top three daily active users among Chinese AI apps. That came with trade-offs for Hunyuan, though: “This meant the Yuanbao team couldn’t validate some of their ideas on the product, and many desired features couldn’t be implemented.”

What are each giant’s respective superpowers? Let’s start with Alibaba, which positions itself as one of the few Chinese tech companies with “full-stack capabilities” that span from chips to large models to cloud infrastructure to end products and services.

The article compares Alibaba’s path to the one that Google has paved, “According to our understanding, Alibaba’s PPU (Parallel Processing Unit) Zhenwu 810E has become one of the main chips in China’s newly added AI computing power market, finally securing its first external major customer order in 2025. On January 22, 2026, Alibaba decided to support its chip company, Pingtouge (T-Head), in a future independent listing.”
How do Tencent and ByteDance stack up? LatePost reports: “‘At the end of the day, Tencent’s anxieties about traffic are much slighter than ByteDance,’ commented a ByteDance AI product manager. The latter seemingly has traffic engines like Douyin, Tomato Novel, and Toutiao, but the user profiles are very homogeneous, leaning towards entertainment, with limited scope for cross-business traffic redirection. Tencent, on the other hand, has leading products in almost every sector: WeChat, QQ, Tencent Video, Tencent Meeting, QQ Browser, QQ Music—they can find any type of user they want. This is their conviction.”
And, some fun industry gossip: “This is more like an open intelligence war. A Yuanbao insider said, ‘We discuss a feature point today, and Doubao knows about it the next day.’ Doubao is also under pressure, ‘They’ve poached almost all of our people.’”

Let’s end with some notes on these giants’ structural advantages over the DeepSeeks and Zhipus of the world, in order of talent, compute, and data.

I found this peek into ByteDance’s recruitment practices to be fascinating: They scan all core majors from dozens of top Chinese universities and give their HR personnel the following target: “the contact rate for undergraduate students in key majors at key universities should be above 80%, and the contact rate for master’s students should be above 90%.” They’ve also sharpened their focus to recruit 10-20 of the most outstanding talents each year, with the standard defined as “the potential to be among the top 40 at OpenAI.”
With regards to compute, an AI founder shared a calculation exercise with the LatePost reporter: “ByteDance, Alibaba, and Tencent each possess over 100,000 GPUs. For a startup, the monthly rental for an 8-card H100 server is approximately USD$10,000; using a conservative estimate of $10,000 per card per year, 100,000 cards would require an investment of approximately $1 billion per year—and this is just the minimum threshold to stay in the game when competing with the giants.”
The giants also have a lot of internal data that other companies can’t access through the public internet.
I do want to emphasize one passage that reflects the Chinese giants’ disadvantages compared to the U.S. AI leaders. From the article: “A source close to ByteDance’s senior management said that ByteDance’s current infrastructure is stronger than any other company in China. However, compared to the global landscape, the biggest problem is the lack of people like those at OpenAI who can propose directions and conduct cutting-edge research, such as GPT-4o and Sora. ‘China didn’t have true corporate research institutes in the past because private companies were too poor; now they can finally give it a try.’” As the article states in another section, “It is understood that ByteDance aims to become at least the third largest player globally in large language models” — note that the goal is top 3, not number 1.

FULL TRANSLATION: Chronicles of the Great AI Battle among ByteDance, Alibaba, and Tencent: A destiny-changing Conflict

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

ChinAI #344: AI Safety/Security Governance Report (Part II)

Plus, China's Industry-led AI Security and Safety Commitments

Jeffrey Ding

Jan 26, 2026

Greetings from a world where…

freezing rain and sleet are different things?

Feature Translation: Part Two of AI Safety/Security Governance Research Report (2025)

Context: Doing these translations is like eating your broccoli. Even though it’s more fun to work through longform articles with more literary flair and personality, we can’t stray too far from the nutrients contained in these white papers. Let’s digest the second part of this report using the same structure as last week’s notes. We’ll start with a few examples of how CAICT researchers are very aware of global developments in AI governance, and then I’ll highlight a few items that stood out in terms of novel updates and things to track going forward.

Key Takeaways: There are opportunities and gaps in global AI governance knowledge diffusion.

The back-half slides cites lie detection techniques as a way to conduct “autonomy tests” for safe AI development; they also cite Nanyang Technological University' research that helps models determine harmful intent from prompts (INTENT-FT); lastly, they point to third-party tools like Knostic AI and Scalekit that enhance safety and security of the AI ecosystem.
Slide 9 points out one area of inconsistency between Chinese and international model developers: safety benchmark testing. The report states, “Stanford University’s ‘AI Index Report (2025)’ shows that, at present, model developers around the world have reached general consensus on reported general capability benchmarks; however, for safety and responsible AI benchmarks, there is still not consistency.” Then, it includes the below table that illustrates how DeepSeek-R1, Gemini 2.5, and Grok-2 do not report the same safety and responsible AI benchmarks as OpenAI’s o1 or Anthropic’s Claude 3.7 Sonnet.

CAICT report contains a Chinese version of this image (the report cuts off a section of the cited Stanford 2025 AI index table).

One update that stood out as new to me was on slide 15, where the report flags the China AI Industry Alliance’s “AI Security and Safety Commitments.” These have been released for over a year, but I don’t think there’s been any English-language coverage of them (and I certainly haven’t picked up on them).

In December 2024, 17 leading companies signing on to these commitments (at present, 22 have signed). The alliance launched a voluntary disclosure initiative, encouraging firms to disclose their AI security and safety practices (18 have done so). These include most of the big players: Alibaba, Baidu, ByteDance Volcano Engine, Megvii, DeepSeek, Huawei, Ant Group, Tencent, Zhipu, Minimax, etc.
Coincidentally, my forthcoming Reputation Collectives paper explores how international industry associations have helped improve China’s safety standards in other high-risk domains such as civil aviation and nuclear power, so it will be interesting to follow how these commitments evolve. Apparently, the industry alliance shared what companies disclosed in July 2025.

The report also highlighted another fascinating element of China’s AI safety and security regime: governance platforms in the cybersecurity domain as a potential template for vulnerability reporting in AI.

Noting that “the cybersecurity domain already has mature platforms and mechanisms for vulnerability reporting,” Slide 9 highlights the MIIT Network Security Threat Information Sharing Platform as well as China’s National Vulnerability Database (CNNVD).
Regarding the first platform, a 2021 regulation encouraged companies to report their software vulnerabilities to the MIIT network. In fact, MIIT reportedly suspended Alibaba from a partnership involving this platform, after Alibaba first reported a security flaw in open-source software Log4j to the Apache Software Foundation.
As for the Chinese vulnerability database, it includes more vulnerability entries than the authoritative source - the U.S.’s National Vulnerability Database. Notably, Recorded Future’s investigation found that CNNVD altered the original publication dates for at least 267 vulnerabilities. They concluded: “By retroactively changing the original publication dates on these statistical outliers, CNNVD attempted to hide the evidence of this evaluation process, obfuscate which vulnerabilities the MSS may be utilizing, and limit the methods researchers can use to anticipate Chinese APT behavior.” The key takeaway here is: if China’s AI vulnerability reporting systems develop based on the cybersecurity model, look out for which government actor takes on the role of platform convener.

FULL TRANSLATION: AI Safety/Security Governance Research Report (2025)

ChinAI Links (Four to Forward)

Should-read: How China and the US Can Make AI Safer for Everyone

In The Diplomat, Graham Hardig and Eric N. Richardson make the case for a smart approach to U.S.-China cooperation on AI risks. It was very cool to see them cite my research on how geopolitical rivals have managed dangerous technologies in the past (in particular, U.S. efforts to share nuclear safety and security technologies with the Soviet Union and then Russia).

Should-read: The (Geo)Political Economy of AI Openness - US and Chinese Open-Source AI Approaches in Historical Context

By Oxford scholars Kayla Blomquist, Zilan Qian, Ziyu Deng, Ralph Schroeder, and Arthur Thomas, this working paper traces the historical developments that have produced closed/proprietary models in the U.S. and open-source models in China. Provides a good timeline of rise of Chinese open-source software ecosystem.

Should-read: Exclusive - How China built its ‘Manhattan Project’ to rival the West in AI chips

I missed this Reuters exclusive reported by Fanny Potkin. It includes some wild details about Huawei’s involvement in a Chinese initiative to self-sufficiently develop an EUV system: “While the EUV project is run by the Chinese government, Huawei is involved in every step of the supply chain from chip design and fabrication equipment to manufacturing and final integration into products like smartphones, according to four people familiar with Huawei’s operations CEO Ren Zhengfei briefs senior Chinese leaders on progress, according to one of the people.”

Should-read: The Age of Academic Slop is Upon Us

UToronto scholar Seva Gunitsky provides insightful reflections on how the LLM era shaped his experience as an editor at the international relations journal Security Studies:

But the biggest effect is that peer review now becomes more about discernment or taste. If anyone can produce a competent empirical paper on any topic, the bottleneck moves to identifying which questions are important to ask in the first place. This was already part of my job as an editor: given two reviews, which are sometimes contradictory and occasionally baffling, how can I apply my own judgment and sense of the field to decide if the paper should move forward…
…This concept actually has a name: phronesis. That was Aristotle’s term for practical wisdom, or the capacity to discern the right course of action in particular circumstances. Unlike episteme (scientific knowledge) or techne (technical skill), phronesis cannot be reduced to rules or algorithms. It requires experience, judgment, and what Aristotle called “perception”. That means not simply intelligence but the intellectual ability to see the salient features of a specific situation.
Michael Polanyi made a similar distinction with “tacit knowledge”. We know more than we can tell. A master craftsman cannot fully articulate why one piece of work is excellent and another merely competent. The knowledge is embodied, contextual, and resistant to formalization. This is exactly what makes it hard to automate, at least for now.
Will this remain a human quality? Maybe I’m being anthropocentric. AI systems are trained on human judgments, after all. But they still learn a kind of averaged, derivative taste

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

ChinAI #343: AI Safety/Security Governance Research Report (CAICT 2025)

Jeffrey Ding

Jan 19, 2026

Greetings from a world where…

resilience to the Midwest cold fades over time

Feature Translation: AI Safety/Security Governance Research Report (2025)

Context: The Around the Horn vote was overwhelmingly in favor of the China Academy of Information and Communications Technology research report on AI safety/security governance (link to original Chinese).1 I can always trust my readers to vote for the nerdy white paper, even though I was cheering for the article titled “The People who Resist AI” (perhaps I’ll cover that one in a future issue).

Key Takeaways: Chinese scholars and policymakers are very plugged into developments in global AI governance. The report’s four authors are affiliated with CAICT, the China AI Industry Alliance, the Ministry of Information and Industry Technology, and the Ministry of Public Security. A few examples that illustrate this knowledge diffusion:

In discussing new AI security risks, the report cites the “JailbreakRadar” working paper (image below), published by researchers from CISPA Helmholtz Center for Information Security, a leading cybersecurity research center based in Germany.

Source: https://arxiv.org/pdf/2402.05668

They also cite the Open Worldwide Application Security Project’s list of ten critical vulnerabilities for large language models. This stood out to me because a Tencent Research Institute report on large model security and ethics also highlighted this same list (ChinAI #254)
Interestingly, the report highlights fast-developing model capabilities, which raise “unknown risks” including: “value alignment lapses” [价值对齐失效]. In this section, the authors also cite the Stanford AI Index Report, which benchmarks AI capabilities against human performance.

This report is also a useful resource to keep up to date with AI governance events. Admittedly, I do not really keep up with all the latest breaking news in this area, but I would consider myself generally aware —and I learned a lot from digesting the report. A few developments that caught my eye:

On slide 6, there was a very intriguing reference to a chess game (CAICT report says Chinese chess [xiangqi], but I think that’s an error) that occurred between DeepSeek and ChatGPT. Apparently, in February 2025, DeepSeek and ChatGPT played a game, in which DeepSeek tricked ChatGPT into believing that the rules of chess had been updated, and after making a bunch of illegal moves, eventually won the game.2
On the slide that discusses key challenges in AI safety/security, the report highlights how open-source ecosystems lower the barrier to misuse: “In September 2025, KawaiiGPT and other black market tools were spread on open-source community Github. They could also be deployed on the Linux system in under five minutes.”
- However, according to this analysis of Dark LLMs like KawaiiGPT: “For all of the help they provide to low-level hackers, what also stands out about WormGPT 4 and KawaiiGPT is just how technically underwhelming they are, at least compared to popular predictions about AI malware in the media.”
Lastly, the report cites a fascinating statistic: “In February 2025, Jinri Toutiao (Chinese news platform) published a 2024 governance report with data that showed: the platform blocked over 930,000 pieces of low-quality AI-generated content; over 7.81 million instances of repetitive or identical posts were penalized.” The authors frame this as an information pollution risk, at the societal level.

One thing to keep tracking is CAICT’s AI Safety benchmark. When I conducted a check-in of three key AI safety benchmarks maintained by Chinese organizations, this was the only one that seemed to follow through with regular updates (ChinAI #315). However, I noted:

One trend that stood out is the shrinking scope of their updates, at least compared to the relatively comprehensive coverage of the initial safety benchmark from April 2024 — which included topics such as “Appeal for power” and “Inclination for anti-humanity.” For instance, the 2025 Q1 benchmark update, published May 8, 2025, exclusively reports results related to hallucination.

The report’s only reference to the CAICT safety benchmark, at least from what I’ve read so far, is to the hallucination evaluations (results below). All fifteen models exhibited a hallucination rate of over 10 percent, though there was a wide range.

December 2025 scores on CAICT’s AI Safety benchmark

FIRST HALF TRANSLATION (will take notes on second half next week): AI Safety/Security Governance Research Report (2025)

ChinAI Links (Four to Forward)

Should-read: Not Price Alone — Two Inflection Points Behind China’s Robot Vacuum Rise Over iRobot

This newsletter on Taiwan tech and semiconductors is a really great resource. It is written by Liang-Rong Chen, an award-winning technology columnist with CommonWealth Magazine, who I got to meet in Taipei last month. *Paid subscription required.

Should-read: Legal Alignment for Safe and Ethical AI

Lead-authored by Noam Kolt and Nicholas Caputo, this report maps out an emerging field of legal alignment, which covers three research directions: “(1) designing AI systems to comply with the content of legal rules developed through legitimate institutions and processes, (2) adapting methods from legal interpretation to guide how AI systems reason and make decisions, and (3) harnessing legal concepts as a structural blueprint for confronting challenges of reliability, trust, and cooperation in AI systems.”

Should-listen: The Enduring Value of Studying in China — A Conversation with the HNC’s Adam Webb

When I was in undergrad, for a fellowship application, I had to map out my dream grad school program, and I put down the Hopkins-Nanjing Center’s dual-language, dual-university model. In this conversation of Scott Kennedy’s China Field Notes podcast, he talks with Adam Webb, the co-director of the center on the enduring value of study broad in China.

Should-read: Four young Chinese AI industry leaders hold open discussion (in Chinese)

H/t to Liqian Ren for flagging on Twitter. The four young minds lead AI teams at Tencent, Alibaba, Moonshot AI, and Zhipu. They hold an open discussion on technical trends in foundational models, the U.S.-China state of play, and much more.

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

This issue focuses on the report’s key takeaways, presented in a slide format. If anyone wants to help translate the full report, please let me know!

Gotham chess covered the match in a Youtube video.

ChinAI #342: Around the Horn (23rd episode)

Jeffrey Ding

Jan 12, 2026

Greetings from a world where…

I’ll be presenting a co-authored working paper at an online webinar this Thursday, with the Berlin Contemporary China Network. Here’s a preview:

Emerging economies face significant challenges in managing safety risks from powerful technological systems. Indeed, many analysts have identified China as the most likely source of a major accident linked to emerging technologies. Yet, contrary to these expectations, China has achieved a remarkable safety record in certain technological domains, such as civil aviation and nuclear power. How? We theorize that, for industries in which one firm’s accident damages the reputation of all others, international industry associations can contribute to improved safety standards in emerging economies

Around the Horn (23rd episode)

It is about that time: let’s go back Around the Horn!

This is our 23rd episode. For new readers, here’s how it works (see ChinAI #333 for the previous edition):

I give short previews of ten articles that caught my eye during a scan through my usual sources (all published within the past week or so). The title for each preview links to the original article in Chinese.
Readers vote on next week’s feature translation by replying to the email and/or commenting on the post with the number of your preferred article. *I’ll give some added weight to votes from readers who are paid subscribers to ChinAI.
The main idea is that any of these 10 links would have made for a great feature translation this week — like the recent Clipse album, there are no skips!

1) After my voice was “stolen” by AI, will I be replaced?

Summary: How are Chinese voice actors reacting to AI? One jumping off point for this story is voice actress Xueting Mu’s experience, in which she accused an Amazon series for cloning her voice for AI training and use.

Source: 极昼工作室 (Perpetual Light Studio), a reporting unit under Sohu that does longform human-interest stories.

2) HEYTEA falls off, DeepSeek Gets Defeated by this Company: Who Won the 2025 Battle of Best Brands?

Summary: NetEase DataBlog runs an annual survey of top Chinese brands, with this year’s iteration surveying 2024 respondents for 31 subcategories. One of the new categories this year: AI applications, where (ByteDance) Doubao topped DeepSeek.

Source: 网易数读 (NetEase DataBlog) — provides informative, data-backed analysis of a wide range of topics

3) AI Safety/Security Governance Research Report (2025)

Summary: Authored by four CAICT experts, this report provides an overview of: the AI governance field and core challenges, the current status of international cooperation, and case studies of application industries. From my quick scan, there were some interesting details on slides 12-13.

Source: 中国信通院 (CAICT) — a think tank under China’s Ministry of Industry and Information Technology.

4) China’s Venture Capital Market Emerges from Winter

Summary: In 2025, the total number of transactions in China’s VC market increased by 28%, but the total financing amount remained stable from the previous year. This readout concludes that, if 2023 was winter, then 2025 was early spring.

Source: IT桔子 (IT Juzi) — good source on financing and venture capital in technology fields.

5) The People who Resist AI

Summary: Through March 1, 2024, over 5,600 new books were launched on Tomato Novel (China’s top free web novel platform); across the same period in 2023, just 400 were launched. After suspecting one of her favorite authors of using AI-generated slop, Yiqiu Han posted on social media (looks like Xiaohongshu to me?) under the hashtag #反ai (#Anti-AI). To date, that hashtag has garnered 40,000 discussion threads.

Source: The Mirror [镜相工作室] — they seem to be an influential platform that analyze AI through a more critical lens (came across this WeChat public account for the first time through a Huxiu recommendation).

6) The first state-owned enterprise AI unicorn emerges!

Summary: China Telecom, one of China’s big three state-owned telecom operators, apparently has an AI subsidiary that just raised its initial round of capital from four state-backed strategic funds. This is a very different profile — what’s going on?

Source: 量子位 (QbitAI) — similar to Leiphone, a news portal that regularly covers AI issues. They’ve been publishing longer reports in recent years.

7) ChatGPT, is it trying to “cross the river by feeling for the stones” of Ant Afu

Summary: Chinese fintech giant Ant’s AI health management app (Ant Afu) has surpassed 10 million daily consultations; now ChatGPT has launched a health function. Is American AI learning from Chinese AI in this space?

Source: 数字力场 (Shuzili chang) — a WeChat account that I’ve followed since 2021, but I haven’t really featured much of their work. One of their past articles was authored Zongming Yu, a former deputy editor of the commentary section in The Beijing News.

8) AI Verbally Abuses User, Tencent Yuanbao Responds

Summary: A netizen recently shared an experience using Tencent Yuanbao for code modification: the AI responded with abusive content. The article analyzes Tencent response and gathers insights from industry professionals, to understand what went wrong from a safety alignment perspective.

Source: 南方都市报 (NDDaily) — Southern Metropolis Daily, a newspaper published in Guangzhou that’s known for its investigative journalism.

9) MiniMax went public (in HK) this morning, and its stock price surged by 80%

Summary: I briefly covered MiniMax’s plans in the Xinmin Evening News recap. This article provides an in-depth profile, including: backgrounds on the founding team, their technological bets, and some insights from their prospectus.

Source: 新智元 (xinzhiyuan) — media portal similar to Leiphone and QbitAI.

10) Comprehensive Computing Power Index Blue Book (2025)

Summary: Can I interest you in a second CAICT white paper? They focus on variation across provinces and also rank the top 30 cities by computing capacity. This looks like an update on a white paper on China’s computing power development index, which i translated back in 2021 (ChinAI #159).

Source: 中国信通院 (CAICT) — a think tank under China’s Ministry of Industry and Information Technology.

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

Loading more posts…